That is a very interesting proposal because I believe it would involve a non-typical NN struture. The main problem is what to do as nodes dissappear (go offline, reboot, logout...) and where in the structure to insert new nodes. How would such a NN get trained? It would have to be done in such a way that the NN could still function as its structure was dynamically changing. This ability to dynamically alter the structure of the NN could be very advantagous and even lead to more efficient NNs
Slashdot Mirror
That is a very interesting proposal because I believe it would involve a non-typical NN struture. The main problem is what to do as nodes dissappear (go offline, reboot, logout...) and where in the structure to insert new nodes. How would such a NN get trained? It would have to be done in such a way that the NN could still function as its structure was dynamically changing. This ability to dynamically alter the structure of the NN could be very advantagous and even lead to more efficient NNs